Observability & SLOs for Inference

Turning a vague sense that requests feel slow into a percentile target you can actually alert on, and tracing which hop is responsible when it's breached.

Part of: Serving Open Models in Production